Open Whisperer gives Claude Code a real voice loop on your Mac: dictate your prompt, and every reply is read back with streaming text-to-speech that starts on the first sentence. Hands-free, 100% on-device, free and open source.
# hold Ctrl and just say it —
you ▸ refactor the auth middleware to use the new token helper
…transcribed on-device, typed into Claude Code, submitted
claude ▸ Done. I moved the verification into verifyToken() and
updated the three call sites. Want me to add a test?
▸ spoken aloud, streaming, starting on the first sentence
you ▸ yes, and run it
Hold a key, or go fully hands-free with "initiate". Your speech is transcribed by on-device Whisper and typed straight into Claude Code — auto-focused and, if you say "send", submitted for you.
A one-click hook reads each response aloud with Kokoro text-to-speech — streaming, so it starts speaking before Claude has finished writing. Say "hold on" to interrupt instantly.
Speech-to-text and text-to-speech both run in-process on the Apple Neural Engine. Your prompts, your code, and your voice never touch a third-party server.
MIT-licensed, no subscription, no account. Works with the Claude Code CLI and the VS Code extension. Read the source and see exactly how the hook works.
Download the DMG and drag Open Whisperer to Applications. On first launch it downloads the on-device Whisper and Kokoro models — no Python, no virtualenv.
Click the one-click setup in the menubar. It adds the hook so Claude's replies are spoken, and auto-focus targets your editor. No config files to hand-edit.
Hold Ctrl and speak, or say "initiate" to go hands-free. Your prompt lands in Claude Code and the reply streams back out loud, sentence by sentence.
Coding with an AI assistant is a loop: describe, read, correct, repeat. Typing every prompt and scanning every reply keeps you pinned to the keyboard. Voice mode breaks that: you can pace, sketch on paper, or lean back and still drive the session — describe the change out loud, listen to what Claude did, and say "yes, run it."
It's also just faster for the messy middle of a task — the "no, use the other helper, and keep the error handling" clarifications that are quicker to say than to type. Because everything runs on-device, none of it depends on a network round-trip beyond Claude's own API.
Prefer the terminal-first workflow? The same setup works for Codex CLI. Curious how it compares to plain macOS Dictation? That page breaks it down.
Free, open source, 100% on-device. macOS on Apple Silicon.